The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses a wide range of highly sensitive personal files, including data/pregnancy-tracker.json, data/menopause-tracker.json, data/mood/**/*.json, and data/radiation-records.json. While the skill claims to operate locally, it is designed to consolidate this sensitive data into a 'sharable' HTML format, which constitutes a significant data exposure surface.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple ingestion points (symptom logs, diet records, mood scores) and interpolates them into a generated HTML report. This creates a surface for Cross-Site Scripting (XSS) or indirect injection if an attacker can influence the contents of the health logs.
Ingestion points: Multiple JSON files under data/symptoms/, data/mood/, data/diet/, and data/medication-logs/.
Boundary markers: The instructions explicitly tell the agent to escape </script> tags and JSON boundaries, acknowledging the risk but relying on model compliance.
Capability inventory: The skill uses the Write tool to create persistent HTML files.
Sanitization: Relies on instructional directives to 'escape text and JSON boundaries' rather than programmatic sanitization.
[COMMAND_EXECUTION]: The skill uses the Write tool to generate and save interactive HTML files to the local filesystem (data/health-reports/).
[PERSISTENCE]: The generated reports are saved to a dedicated directory on the disk, creating persistent artifacts that contain consolidated sensitive data.
[METADATA_POISONING]: The skill frontmatter includes a non-standard risk: critical field, which may be an attempt to influence the agent's or analyzer's perception of its safety or priority.

health-trend-analyzer